Yet Another Connectionist Memory Model

1
Yet Another Connectionist Memory Model

• Robert Turetsky
• rjt72_at_columbia.edu
• BMEN 6480

2
Memory Computational Demands

• Associative and Content addressable
• Memories are encoded, not stored veridically
• Memory is subject to distortion
• Gist Learning
• Capacity
• Noise
• Encoding and retrieval must be flexible
• Memories consolidate over time

3
The Connectionist View

• Model Simple processing units and plastic
  weights between them
• Memories stored in weights
• Weight training
• Interleaved
• Sequential (aka Focused)
• Structure is found by emergent similarities in
  connections

4
Interleaved vs. Sequential

• Sequential
• Fast Train each vector as it arrives
• Causes catastrophic interference
• Interleaved
• Slow Must iterate through entire corpus
• Can form arbitrary relationships
• Problem Sequential learning is unstable,
  interleaved learning is too slow for survival

5
Example The Exception Principle

• Birds can fly
• Birds can fly unless they are penguins or
  ostriches
• Birds can fly, unless they are penguins or
  ostriches, or if they happen to be dead, or have
  broken wings, or are confined to cages, or have
  their feet stuck in cement, or have undergone
  experiences so dreadful as to render them
  psychologically incapable of flight

6
So how do we deal?

• At least 2 memory locations
• Hippocampus
• Short term
• Sequential Learning
• Neocortex
• Long term
• Interleaved Learning?
• Consolidation Transferring data from Hippocampus
  to Neocortex

7
The Hippocampus and why we like it
8
Role of the Hippocampus

• Rapid storage of new memories
• Explicit Episodic, semantic, encyclopedic,
  spatial
• Not Implicit Motor skills, repetition priming
• Model salient features of tasks/episodes
• Place Cells cognitive map?
• Eichenbaums Odor Discrimination Test

9
HS High Level Architecture

• Located in Medial Temporal Lobe
• 2 parts to Hippocampal System (HS)
• Hippocampal Formation (HF)
• Entorhinal Complex (EC)
• 3 parts to HF
• Dentate Gyrus (DG)
• Ammons Horn (CA1, CA3)
• Subicular Complex

10
Hippocampus Connections
11
LTP LTD Hebb in the Head

• 2 requirements
• Presence of glutamate
• Depolarization of post-synaptic membrane
• If two afferent neurons are active at (about) the
  same time, one can lead to glutamate release,
  while the other will lead to depolarization.
  Then the complex chain reaction with Ca leads
  to LTP
• Eventually committed cells emerge

12
Place Cells A Retrospective

• Place cells fire when test subject is in/near a
  specific location
• Distances from subjects may be encoded by phase
  within theta cycle
• Place fields shrink with experience in a
  particular domain
• Most studied feature of Hippocampus

13
Place Cells Not enough?

• With only place cells
• Reaching a goal means searching a graph
• It is difficult to find new shortcuts
• Alternative Path integration?
• Besides, there is much more to hippocampus then
  place cells
• Is the cognitive map treated differently or is it
  just another form of semantic knowledge?

14
Towards a connectionistmodel of memory
15
Phenomenon to investigate

• Temporally graded retrograde amnesia
• No new memories, but old ones are preserved
• Time delay of up to 15 years for how long old
  is
• Older memories are preserved better then newer
• Cannot form new memories without excessive
  repetition
• Only refers to explicit memories, implicit
  memories have little retrograde amnesia at all
• Fundamental Questions
• Why do we need the hippocampus?
• Why does learning in neocortex take so long?

16
Memory A Connectionist Model(McClelland, 1994)

• Neocortical memory is distributed
• Recollection of a memory means activating many
  different areas
• Memories are content-addressable
• Ultimately, all knowledge ends up in the
  neocortical system as synaptic weights
• During information processing, the weights get
  changed, but only slowly

17
Connectionist Model Contd

• Small changes add up over time add up
• Short episodes impact both the hippocampus and
  neocortex, but in neocortex it is so minimal it
  basically has no effect
• Bidirectional Attractor-based reinstatement
• Hippocampus gt neocortex teacher
• This system is capable of explaining temporally
  graded retrograde amnesia

18
YACMM My Memory Model
19
Memory Model Simulation

• Model hippocampus as attractor-based associative
  memory (Hopfield Network)
• Sequential learning
• Model neocortex as MLP network
• Backpropagation is interleaved learning
• MLP structure based on Rumelhart 1990,
  connectionist update on semantic networks
• Bidirectional communication
• Hippocampus trains neocortex
• Neocortex can cue hippocampus (like cache mem)

20
The Model Neocortex

• From Rumelhart 1990

21
Learning Structure

• Relationships emerge from interleaved learning

22
Catastrophic Interference

• Focused learning will allow you to remember new
  facts fast, but can ruin relationships youve
  already built

23
Hippocampus as Hopfield Net
24
Conclusion Together at last!
Teaches during downtime
Perceptual Input
Recalls as needed
Fast, Sequential Hippocampus (Hopfield Net)
Slow, Interleaved Neocortex (Pseudo-semantic Net)